CM: Data Analysis workflow for continuous monitoring data from wearables, eHealth devices and ICU monitors

The rapidly growing availability of continuous monitoring data derived from sensors integrated in wearables/implantables or in monitoring systems, such as in ICU, offers unseen opportunities for learning of mechanisms of biomedical processes from longitudinal data. In contrast to large stationary or time-spot related data assessment, continuous monitoring allows easy assessment of causality due to time-course analysis.

However, the analysis of continuous monitoring data is hampered by data quality issues as well as the mere fact, that even sophisticated monitoring devices can monitor only a small subset of medical parameters which might be involved in the process under consideration. Moreover, the required analysis strategies for monitoring data will differ significantly between monitoring a few parameters assessed on tight sampling rates within weeks, e.g. ecg, eeg or glucose data, and monitoring a multitude of parameters along hours or a few days with low sampling rates as is the case in ICU. In addition, the parameters assessed by continuous monitoring are rarely directly associated with the biological mechanisms driving disease propagation or therapy, but they serve primarily as correlates. Hence, biological interpretation of the data requires the inference of the underlying model from the data.

Data quality issues to be tackled in continuous monitoring data analysis are primarily

  • Outlier analysis to identify sensor errors and sensor drifts
  • Merging heterogeneous time series with highly different sampling rates
  • Imputation of missing values

Data analysis issues are primarily driven by the need for model identification, whereas the model can be either mechanistic (e.g. heartbeat models), correlative (e.g. fourier series assuming a linear response model) or data driven (e.g. identification of variability-related patterns in order to quantify approaching a critical point, based on phase transition theories). Recently promising methods have been reported from a combination of Takens Imbedding and Deep Learning. Each of these data analysis methods have their strengths and weaknesses, where no method can claim superiority in all cases of applications, such that a rule-based decision support system for optimal monitoring data assessment is required.